Abstract
In this paper, we propose a method for raindrop removal from rain-drop coated images acquired during nighttime rainy weather. To the best of our knowledge, no previous deep learning-based method for raindrop removal has been developed for nighttime images. This is because it is difficult to identify raindrop regions in images acquired at night, making it difficult to remove raindrops, and there is no nighttime raindrop image dataset to train a deep learning neural network.To solve these problems, we propose a method to create nighttime raindrop images from daytime images without raindrops by using a deep learning based semantic segmentation method and a method used in the field of computer graphics. In order to improve the ability to identify raindrop regions, we propose a GAN architecture that incorporates the U-shaped Transformer structure into the GAN generator. In addition, we propose to apply histogram flattening to the input images as a preprocessing step during training so that the GAN can be trained stably on a nighttime raindrop image dataset. Experimental results are presented to verify the effectiveness of the proposed method.
